Capability Gap

All companies

Scale AI

Took $14.3B from Meta for 49% of itself and lost most of its frontier-lab book within weeks — the cleanest natural experiment in why neutrality is the product.

high confidence5 minupdated 2026-08-29ai labs · data · concentration
Vertical
Expert data for frontier labs
Founded
2016 (Y Combinator)
Headquarters
San Francisco
Raised
~$1.6B equity, plus Meta's $14.3B for a 49% non-voting stake
Last valuation
$29B (June 2025, on the Meta investment); $13.8B in 2023
Revenue
$870M in 2024; ~$2B annualised 2025 per one estimate; company guiding to 'surpass $1B' in 2026 — the bases are incompatible and the real figure is unknown. Gross margin '50%+'.
Status
Active but destabilised; customers fled post-Meta; 200 FTEs and ~500 contractors cut July 2025; pivoted to defence and enterprise

Scale AI is the cautionary tale of the expert data vertical, and its lesson is unusually precise: when your buyers compete with each other, an equity relationship with one of them is an exit event for the rest of your revenue.

Founded in 2016 by Alexandr Wang and Lucy Guo through Y Combinator, Scale climbed a conventional valuation ladder — $1B (2019, Founders Fund) → $7B (2021, Greenoaks) → ~$13B (2024, Accel) → $14B (2024, Amazon and Meta) → $29B on Meta's June 2025 investment (Wikipedia; Sacra).

The deal

In June 2025 Meta paid $14.3B for a 49% non-voting stake — Wikipedia gives $14.8B and the two figures circulate interchangeably — and Wang left to run Meta Superintelligence Labs. Scale nominally stayed independent (Wikipedia; Computerworld). The structure — a large minority stake plus an acqui-hire, deliberately short of control — is what a buyer does to avoid merger review. It did not avoid the commercial consequence.

What it did to the rest of the book

CustomerWhat happened
GoogleCut ties. Had spent ~$150M with Scale in 2024 and planned ~$200M in 2025
OpenAIDeparted, having already been reducing
MicrosoftPulled back
xAIBegan exploring alternatives
MetaBy August 2025 was routing work to Surge AI and Mercor instead

Sources: Sacra, Computerworld, TechCrunch.

In July 2025 Scale cut 200 FTEs (~14% of staff) and terminated around 500 contractor relationships, concentrated in data labelling (Sacra). By August 2025, TBD Labs researchers inside Meta reportedly considered Scale's data low quality and preferred competitors — so the buyer that had just paid $14.3B for half the company was itself buying elsewhere. Vertical integration failed on quality grounds, which is the part most retellings omit.

The analyst read, from Thomas Randall of Info-Tech: Meta's move signalled "a trend toward vertical integration and supplier lock: owning the data annotation pipeline to secure control over quality, provenance, and scalability" (Computerworld). It is also why Google structured its approach to Mechanize as a licence-plus-hire rather than an acquisition. See One customer is a binary event.

Revenue, which nobody can pin down

Gap in the record

Scale's actual 2025 and 2026 revenue could not be established. One source gives ~$870M for 2024 [WEAK]; Sacra says $1.5B annualised exiting 2024 and ~$2B estimated for 2025; third-party reports say Scale expects to surpass $1B in 2026 with enterprise at ~$200M annualised [WEAK]. If the $2B-2025 and $1B-2026 figures were on the same basis, revenue roughly halved — but they almost certainly are not (gross vs net, bookings vs revenue). Do not compute a growth rate from these numbers.

Scale's own 2025 retrospective claims over $1B in new business — bookings, not revenue — with nearly half booked in Q4, the data business profitable in H2 2025, and headcount up 500+ (Scale blog). Gross margin is put at "50%+" (Sacra), which if accurate would be the highest in the private cohort and well above Mercor's leaked 33% — but it is a single secondary source. See GMV is not revenue.

The pivot is real and away from frontier labs: Project Thunderforge (~$500M defence), Golden Dome, a $99M Army contract, two Department of War awards totalling ~$200M, plus Mayo Clinic, BP and Allianz as enterprise logos, robotics (150,000+ data hours) and SEAL evaluations.

The supply side, and its collapse

Remotasks (2017) handled computer-vision and autonomous-vehicle labelling in South-East Asia and Africa. A 2022 Oxford Internet Institute study found it met minimum fair-work standards on 1 of 10 criteria, with pay below one cent per task in some regions; it was abruptly shut in Kenya, Nigeria and Pakistan in early 2024 and had obscured its Scale affiliation (Wikipedia).

Outlier is the higher-skill LLM arm. Baseline rates there reportedly fell from ~$40–50/hr to ~$15–20/hr by late 2025, with ~40 hours of unpaid onboarding common and the Dallas generalist team eliminated in October 2025 (Breaking Even) [WEAK — community-data analysis, not press].

Quality failures, documented

This is the best public evidence anywhere that vendor QC in this vertical is weak. Inc obtained internal logs showing Scale's "Bulba Experts" programme for Google's Bard and Gemini was plagued by spam for 11 months, March 2023 to April 2024: contributors using ChatGPT despite an explicit prohibition, writing gibberish, faking required advanced degrees, non-native speakers on English-fluency projects, accounts logged 18+ hours daily implying shared credentials, and one incident in which "the Allocations Department had dumped 800 spammers into our team." Scale says safeguards caught spam before delivery (Inc). See Who is actually on the other end.

Litigation and leadership

Three worker suits between December 2024 and January 2025 alleged wage theft, misclassification and psychological harm from disturbing content; one plaintiff said he was promised $25/hr and received a fraction (Wikipedia; Inc). The Department of Labor opened an investigation in March 2025 and dropped it in May 2025 (TechCrunch).

Gap in the record

Leadership is unsettled in the record. Jason Droege became CEO in June 2025; a separate compilation reports Francis deSouza named CEO in July 2026, giving three CEOs in fourteen months. The two accounts were not reconciled and the current CEO should be confirmed before citing.